•STDnet-ST is a novel spatio-temporal ConvNet for small object detection.•STDnet-ST exploits the correlation of promising regions between frames.•An efficient tubelet linking is performed to link ...small objects across video frames.•A novel tubelet suppression algorithm is proposed to avoid unprofitable tubelets.•STDnet-ST outperforms its state-of-the-art counterparts for small target detection.
Object detection through convolutional neural networks is reaching unprecedented levels of precision. However, a detailed analysis of the results shows that the accuracy in the detection of small objects is still far from being satisfactory. A recent trend that will likely improve the overall object detection success is to use the spatial information operating alongside temporal video information. This paper introduces STDnet-ST, an end-to-end spatio-temporal convolutional neural network for small object detection in video. We define small as those objects under 16×16 px, where the features become less distinctive. STDnet-ST is an architecture that detects small objects over time and correlates pairs of the top-ranked regions with the highest likelihood of containing those small objects. This permits to link the small objects across the time as tubelets. Furthermore, we propose a procedure to dismiss unprofitable object links in order to provide high quality tubelets, increasing the accuracy. STDnet-ST is evaluated on the publicly accessible USC-GRAD-STDdb, UAVDT and VisDrone2019-VID video datasets, where it achieves state-of-the-art results for small objects.
Full text
Available for:
GEOZS, IJS, IMTLJ, KILJ, KISLJ, NLZOH, NUK, OILJ, PNG, SAZU, SBCE, SBJE, UILJ, UL, UM, UPCLJ, UPUK, ZAGLJ, ZRSKP
The inversion of subsurface reservoir properties is of profound significance to the oil and gas energy development and utilization. The strong heterogeneity and complex pore structure of underground ...reservoirs pose a challenge to efficient oil and gas energy development and utilization across the world, which increases the necessity of developing an efficient reservoir properties inversion method. However, traditional model-driven methods are confronted with the challenges of strong nonlinearity and geological heterogeneity. Moreover, previous studies rarely emphasized the importance of nonlinear feature selection and transfer learning (TL). Aiming to address the research gaps, a reservoir properties inversion method was proposed by combining random forest (RF) feature selection, bidirectional temporal convolutional network (BiTCN), bidirectional gated recurrent units (BiGRU) network, multi-head attention (MHA) mechanism and TL strategy. First, the RF was used to screen the features with significant correlation with the target reservoir properties. Thereafter, combining BiTCN and BiGRU network to leverage their complementary strengths, a parallel dual branch feature learning network was constructed to learn richer geological information from logging data. Meanwhile, MHA was introduced to fuse the output features of the dual network structure. Finally, the fused features were passed through the fully connected module to output the inversion results. TL was used to associate the correlation between reservoir properties and model inversion to improve the inversion performance. The application results with actual field data showed that the proposed method was accurate and robust in reservoir properties inversion. This study can provide a new way for reliable reservoir properties inversion and promote the application of artificial intelligence in data-driven energy science.
•A parallel dual branch network for reservoir properties inversion is newly proposed.•A feature selection combining correlation coefficient and random forest is developed.•Transfer learning contributes significantly to improve the inversion accuracy.•Actual data application results show that the method is feasible and effective.
Full text
Available for:
GEOZS, IJS, IMTLJ, KILJ, KISLJ, NLZOH, NUK, OILJ, PNG, SAZU, SBCE, SBJE, UILJ, UL, UM, UPCLJ, UPUK, ZAGLJ, ZRSKP
Lithium-ion batteries (LIBs) need to maintain high energy efficiency and power level in several application scenario. Accurate state of health (SOH) forecast is essential for designing a safe and ...reliable battery management systems (BMS). Temporal convolutional network (TCN) is a prevailing deep learning method for estimating the SOH of lithium-ion batteries. However, the hyperparameters in the network are usually difficult to predefine, which poses a challenge for the SOH estimation accuracy in real-world. To solve this problem, this paper proposes a data-driven estimation approach, where the TCN is combined with the modified flower pollination algorithm (MFPA) to determine the network topology. After hyperparameter optimization, the external sensor raw data and identified ohmic resistances trajectories in the equivalent circuits model (ECM) are both input to the TCN model to estimate SOH of LIBs. In contrast to prior approaches for feature extraction, this method is not only conductive to improve SOH estimation accuracy, but also can reduce on-board estimation computing burden. We carry out experiments on the same type of cells from NASA public data resources. The experimental results systematically validate the superiority of the proposed method, which covers high estimation accuracy, great robustness to varied training set and satisfied universality to different batteries.
Display omitted
•The MFPA algorithm is introduced to optimize several key hyper parameters in the TCN structure•Extracting external morphological features from the raw voltage and current curves•Ohmic resistances trajectories in the ECM with aging mechanisms to improve the SOH estimation accuracy•High estimation accuracy, great robustness to varied training set and satisfied universality to different battery types
Full text
Available for:
GEOZS, IJS, IMTLJ, KILJ, KISLJ, NLZOH, NUK, OILJ, PNG, SAZU, SBCE, SBJE, UILJ, UL, UM, UPCLJ, UPUK, ZAGLJ, ZRSKP
Power system condition prediction is to predict key parameters such as voltage and current of its main components to ensure the reliability and safety of electrical system operation. However, as ...satellite missions and operating conditions become more complex, the state of electrical systems is affected by satellite loads and platforms. Therefore, it is difficult for offline-trained state prediction models to dynamically adapt to changing operating conditions and changing relationships between parameters. To overcome these problems, an online deep learning method, On-line Rules-limited Temporal Convolutional Network (O-R-TCN ), is proposed to improve the adaptability of the state prediction method. Firstly, a temporal convolutional network (TCN) online weight update method based on a recursive extreme learning machine is proposed, which realizes the self-learning of dynamic parameter relations under dynamic operation conditions to obtain higher prediction accuracy. Moreover, considering that anomalies within the monitored data may lead to undesired model updating, an association rule mining approach is proposed based on the Frequent Pattern-Growth (FP-Growth) fusion with trend sign aggregation approximation (TSAX). Avoid model updates caused by system exceptions or failures. Experiments are implemented based on the actual satellite power system telemetry data. The results illustrate that the proposed method can self-learn the varying relationship between the parameters and update the state prediction model during the working state change. At the same time, abnormal data can be effectively detected by the association rule, which ensures the effectiveness of the proposed method.
With the high percentage access of photovoltaic (PV) power generation, accurate and stable short-term PV power generation forecasting has become popular to the existing power system planning and ...operation. This paper proposes an ensemble learning method based on signal decomposition, deep learning, and optimization strategy for forecasting short-term PV power. At first, the original PV series is decomposed by utilizing the complete ensemble empirical mode decomposition with adaptive noises (CEEMDAN). Then, the decomposed PV series are separately allocated to different deep temporal convolutional networks (DeepTCNs) for forecasting. Finally, the multi-verse optimizer strategy based on no-negative constraint theory (NNCT) is introduced to integrate the weight coefficients of the ensemble DeepTCNs strategy and reconstruct eventual forecasting results. The case studies on real-time PV data from Alice Springs, Australia, present that the proposed method is superior to other benchmark methods in four conventional performance indexes and two statistical tests, demonstrating the validity of the proposed method in forecasting PV power.
Recently, global attention has been paid to climate change. On this account, the market-based carbon pricing scheme is developed to limit greenhouse gas emissions, where a proper grasp of the pricing ...mechanism is crucial for alleviating global warming. Accordingly, we propose a novel method to interpret carbon price dynamics, concurrently deriving the precise prediction and causality. Due to the nonlinearity and nonstationarity of carbon prices, we develop a real-time decomposition approach coupling the multiple ensemble patch transform (MEPT) and the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN). The MEPT captures the multi-resolution trends of the carbon prices series exactly, and then the ICEEMDAN extracts the fluctuation patterns. Additionally, we collect the numerous potential factors, involving energy sources, energy prices, stock market indices, and economic information. Furthermore, we developed causal temporal convolutional networks (CTCNs) to realize the accurate prediction and the proper causal inference simultaneously. The experimental results on the European Union Allowance (EUA) confirm the effectiveness and necessity of the real-time MEPT-ICEEMDAN decomposition. Moreover, the proposed MEPT-ICEEMDAN-CTCN model exhibits significant superiority in multi-step-ahead and quantile forecast, which realizes the 0.73881%, 1.04461%, and 1.23495% MAPE in one-, five-, and ten-step-ahead forecast respectively and 0.00032 PDQ0.1 and the 0.00285 PDQ0.9 in the quantile forecast. Meanwhile, it reveals the nonlinear Granger causality across the various horizons and quantiles for the first time. It is instructive and inspiring for policymakers, carbon-consumed industries, investors, and researchers.
•The Granger forecast model yields prediction and causality concurrently.•The real-time decomposition method extracts the essential features in practice.•The causal temporal convolutional networks are explainable.•The proposed model achieves the most accurate and stable forecast results.•The results show the differences of causality at various carbon price quantiles.
Full text
Available for:
GEOZS, IJS, IMTLJ, KILJ, KISLJ, NLZOH, NUK, OILJ, PNG, SAZU, SBCE, SBJE, UILJ, UL, UM, UPCLJ, UPUK, ZAGLJ, ZRSKP
Wireless body area network (WBAN) has become a promising technology, which can be widely applied in health monitoring, and so on. However, the performance of a practical WBAN may severely suffer from ...the degradation caused by dynamic nature of wireless channels with the movements of human body. Traditional communication frameworks cannot catch up with the channel variation of dynamic WBANs, which may severely degrade the performance, so an accurate channel prediction model is necessary for developing an efficient transmission strategy. In this paper, we propose a DeepBAN communication framework for dynamic WBANs. In our proposed framework, a temporal convolution network (TCN) based deep learning approach is adopted for channel prediction, the computationally intensive task of which is processed by mobile edge computing (MEC), to reduce the response time. Given the predicted channel conditions, we propose a joint power control, time-slot allocation, and relay selection algorithm to maximize the energy efficiency of the system, taking into account the transmission reliability and end-to-end latency requirements. We evaluate the performance of DeepBAN, and the results show that it can achieve energy-efficient, reliable, and low-latency data transmission in dynamic WBANs, which can improve the system energy efficiency by 15% compared with the stochastic scheduling scheme.
Accurate solar irradiance prediction is crucial for harnessing solar energy resources. However, the pattern of irradiance sequence is intricate due to its nonlinear and non-stationary ...characteristics. In this paper, a deep hybrid model based on encoder–decoder is proposed to cope with the complex pattern for hourly irradiance forecasting. The hybrid deep model integrates complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), encoder–decoder module, and dynamic error compensation (DEC) architecture. The CEEMDAN is implemented to reduce the nonlinear and non-stationarity of the irradiance sequence. The encoder–decoder integrates temporal convolutional networks (TCN), long short-term memory networks (LSTM), and multi-layer perceptron (MLP) for temporal features extraction and multi-step prediction. The DEC architecture dynamically updates the model based on adjacent error information to mine the predictable components of error information. Furthermore, a new loss function is further proposed for multi-objective optimization to balance the performance of multi-step forecasting. In the hourly irradiance forecasting experiments on the three public datasets, the root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) of the proposed model are observed to be in a range of 30.693-34.433 W/m2, 19.398-22.900 W/m2, and 0.9872-0.9902, respectively. Compared to the benchmark models (including MLP, LSTM, and TCN), the RMSE and MAE reduce by 10.76%–22.00% and 5.47%–20.40%, respectively. The experimental results indicate that the proposed model shows accurate and robust forecasting performance and is a reliable alternative to hourly irradiance forecasting.
•Propose a novel hybrid model for hourly solar irradiance forecasting.•Reconstruct temporal features to reduce non-stationarity of irradiance series.•Introduce a dynamic compensation architecture to reduce forecasting error.•Balance multi-step forecasting performance by multi-objective optimization.•The RMSE and MAE reduce by 10.76%–22.00% and 5.47%–20.40%, respectively.
Full text
Available for:
GEOZS, IJS, IMTLJ, KILJ, KISLJ, NLZOH, NUK, OILJ, PNG, SAZU, SBCE, SBJE, UL, UM, UPUK, ZAGLJ, ZRSKP
Accurate heating load prediction is essential to supply-demand collaboration in district heating systems and energy sustainability. With the development of the Internet of Things technology, massive ...monitoring data is collected that provides abundant data sources for the prediction models. However, the missing data from equipment anomalies, network transmissions and the complexity of heat load changing cause difficulty in realizing accurate heat load prediction. Therefore, the hybrid model based on bidirectional long short-term memory network and temporal convolutional network proposed in this paper to improve the accuracy of the heat load prediction from the aspects of reconstruction of missing values and extraction of complex features. The bidirectional long short-term memory algorithm imputes missing values based on bi-directional features of existing data and provides solid data foundation for subsequent training of predictive models. The temporal convolutional network achieves high-level feature extraction and parallel computation, which contributes to the effective modeling of complex changes in heat load and the realization of accurate heat load prediction. Comparative study was conducted with four heat exchange stations of actual environment to evaluate the performance of the prediction model. The results showed the mean absolute percentage error of the hybrid model decreased to 2.47% which is lower than the State-of-the-Art models, such as long short-term memory, temporal convolutional network, and support vector regression etc. The superior results further validated the feasibility of this hybrid model in heat load prediction.
•The missing values in heating system were reconstructed by deep learning model.•A hybrid heating load prediction model was proposed considering missing values.•Detailed experiments were conducted to validate the model's superiority.
Full text
Available for:
GEOZS, IJS, IMTLJ, KILJ, KISLJ, NLZOH, NUK, OILJ, PNG, SAZU, SBCE, SBJE, UILJ, UL, UM, UPCLJ, UPUK, ZAGLJ, ZRSKP
Short text classification is a challenging task in natural language processing. Existing traditional methods using external knowledge to deal with the sparsity and ambiguity of short texts have ...achieved good results, but accuracy still needs to be improved because they ignore the context-relevant features. Deep learning methods based on RNN or CNN are hence becoming more and more popular in short text classification. However, RNN based methods cannot perform well in the parallelization which causes the lower efficiency, while CNN based methods ignore sequences and relationships between words, which causes the poorer effectiveness. Motivated by this, we propose a novel short text classification approach combining Context-Relevant Features with multi-stage Attention model based on Temporal Convolutional Network (TCN) and CNN, called CRFA. In our approach, we firstly use Probase as external knowledge to enrich the semantic representation for the solution to the data sparsity and ambiguity of short texts. Secondly, we design a multi-stage attention model based on TCN and CNN, where TCN is introduced to improve the parallelization of the proposed model for higher efficiency, and discriminative features are obtained at each stage through the fusion of attention and different-level CNN for a higher accuracy. Specifically, TCN is adopted to capture context-related features at word and concept levels, and meanwhile, in order to measure the importance of features, Word-level TCN (WTCN) based attention, Concept-level TCN (CTCN) based attention and different-level CNN are used at each stage to focus on the information of more important features. Finally, experimental studies demonstrate the effectiveness and efficiency of our approach in the short text classification compared to several well-known short text classification approaches based on CNN and RNN.
Display omitted
•Our CRFA approach is composed of three layers.•In the representation layer, the multi-stage attention model is used.•To obtain word and concept features, we utilize different levels CNN and TCN.
Full text
Available for:
GEOZS, IJS, IMTLJ, KILJ, KISLJ, NLZOH, NUK, OILJ, PNG, SAZU, SBCE, SBJE, UILJ, UL, UM, UPCLJ, UPUK, ZAGLJ, ZRSKP
PDF
